Blog
-
Your AI agent had read-only access. Why did it post your data somewhere else?
By
Read-only access limits what an agent takes in, not where it sends it. What the Grok Bot CFO incident shows about outbound permissions and per-action confirmation.
-
Can Users See, Scope, and Revoke What AI Agents Hold? A Study of 1,634 MCP Servers
By
We read the READMEs of 1,634 MCP servers. 85% of credentialed servers use an automated pipe, 8% document single-agent revocation, half document no controls.
-
Can you prove a human approved an AI agent’s access without trusting the vendor?
By
A vendor’s consent log is a statement about its own behavior. AgentAdmit’s hosted ceremony makes the user’s passkey sign a commitment to the exact grant, verifiable offline by anyone holding the export. Here is what it proves, what it does not, and the three honest evidence tiers.
-
Can an AI agent’s grant carry the app’s declared purpose and the user’s own intent, and are they enforced?
By
An AgentAdmit grant records the app’s declared purpose and the user’s own intent. Neither is enforced; scopes decide. Both ride every audit row, alert, and signed ceremony for human review.
-
Can an AI agent be required to get human confirmation for one action, even inside a valid grant?
By
Yes. Confirm each time lets an app owner mark a scope so every call using it is refused until the person confirms that exact call with a passkey on a page the agent cannot complete. Shipped Sep 15, 2026.
-
What do AI agent authorization buyer’s guides leave out?
By
AI agent authorization buyer’s guides evaluate infrastructure tools. The ones we have read skip the same layer: the user consent surface. Here is what is missing and why it matters.
-
Can an AI agent accumulate permissions over time without the user knowing?
By
AI agents accumulate permissions over time in the typical deployment today. Here is why scopes tend to grow but rarely shrink, why token expiry alone does not fix it, and what user-controlled authorization changes.
-
Should an AI agent ever hold its own authority? What the industry and the research now agree on
By
An open letter signed by 100+ organizations, the MCP roadmap, and eight new papers converge: authority cannot live inside the agent. Here is the consensus, and how AgentAdmit ships it.
-
Why do AI agent attacks keep going for the credential, not the model?
By
Major AI agent attacks keep targeting OAuth tokens, API keys, and standing permissions. Here is the four-step pattern and the architectural fix.
-
Prompt Injection Is an Authorization Problem, Not Just a Model Problem
By
Prompt injection cannot be fully solved at the model layer. Authorization is what limits the damage: scope is the control that still holds when the model is fooled.
Get new posts by email
One short email when a new post goes live. No spam, unsubscribe anytime.